Novel Machine Learning Identifies 5 Asthma Phenotypes Using Cluster Analysis of Real-World Data

Area of Science:

• Pulmonary Medicine
• Data Science
• Computational Biology

Background:

• Accurate asthma classification is crucial for tailoring treatments and improving patient outcomes.
• Heterogeneity in asthma necessitates advanced analytical methods for subphenotyping.

Purpose of the Study:

• To leverage machine learning (ML) and large-scale real-world data to identify distinct asthma patient groups.
• To explore asthma heterogeneity and uncover novel subphenotypes.

Main Methods:

• Utilized electronic health records from 13,498 asthma patients (2010-2021).
• Employed k-prototype unsupervised ML for clustering based on clinical and demographic factors.
• Selected optimal clusters using elbow and silhouette plots, validated with LightGBM supervised ML.

Main Results:

• Identified 5 stable asthma clusters with distinct characteristics.
• Cluster 5 represented severe asthma with high blood eosinophilia and poor lung function (FEV1 68.08%).
• Other clusters included young nonsevere asthma, obese patients, normal lung function, and those with poor bronchodilator response.

Conclusions:

• Unsupervised ML effectively classified asthma into 5 clinically significant subphenotypes.
• These identified subphenotypes provide a basis for developing targeted asthma management strategies.
• The findings support personalized medicine approaches in asthma care.